Figure: Overview of GigaWorld-Policy-0.5, an MoT-based action-centered World Action Model. The model consists of a visual expert and an action expert: the visual expert specializes in processing video tokens, while the action expert focuses on action-token modeling. The two experts are connected through multi-modal self-attention, which follows the same causal masking strategy as GigaWorld-Policy to preserve action-centered dependency modeling.
Abstract
World Action Models (WAMs) improve robot policy learning by jointly modeling actions and future visual observations, using future scene evolution as dense supervision for physically grounded action generation. However, a common design in existing WAMs is to explicitly generate future videos at inference time, incurring substantial computational overhead and hindering real-time closed-loop deployment. GigaWorld-Policy addresses this issue with an action-centered formulation, where future visual dynamics are used during training while action-only decoding is used at inference time. Building upon this framework, we present GigaWorld-Policy-0.5, an enhanced action-centered WAM designed for more efficient robot control. During pretraining, GigaWorld-Policy-0.5 adopts a mixed Action-Conditioned World Modeling (AC-WM) and WAM training strategy. This strengthens the coupling between visual dynamics and robot actions and improves the transferability of action representations for downstream policy learning. For efficient inference, GigaWorld-Policy-0.5 introduces a Mixture-of-Transformers architecture that separates visual dynamics modeling and action generation into specialized experts, reducing active computation during action-only inference and achieving 85 ms inference latency on a local RTX 4090 setup. In addition, we employ an agent-based AutoResearch pipeline to systematically search training configurations, enabling more efficient identification of optimal experimental setups while reducing the time and manual intervention required for hyperparameter tuning. Experiments and ablations show that GigaWorld-Policy-0.5 preserves the training benefits of future visual dynamics while improving inference efficiency for robot control.
Features Compared to GigaWorld-Policy
01
Inference Efficiency
Local RTX 409085 ms
C++ deployment latency
The Mixture-of-Transformers architecture separates visual dynamics modeling from action generation, enabling a lightweight action-only pathway for low-latency deployment.
Table: Comparison of inference efficiency and real-robot performance.
Method
Latency on A100 (ms) ↓
Latency on RTX 4090 (ms) ↓
SR on Real-Robot ↑
π0.5
225
110
0.76
Motus
3231
–
0.80
FastWAM
229
182
0.78
GigaWorld-Policy
360
293
0.80
GigaWorld-Policy-0.5
189
110
0.85
+ C++ deployment
140
85
0.85
Figure: Comparison of GigaWorld-Policy-0.5 with baselines on inference latency and success rate across hardware platforms in real-world settings.
02
Mixed AC-WM and WAM Pre-training
Mixed AC-WM and WAM pre-training strengthens action-dynamics coupling and improves the transferability of action representations.
Figure: Success rates at different training steps in the AC-WM ablation study.
03
AutoResearch-Driven Hyperparameter Study
AutoResearch automates pilot runs, metric monitoring, and candidate selection to derive a reliable training recipe with less manual tuning.
Figure: AutoResearch hyperparameter search and training progression on the pick the fruit task. Left: 1K-step pilot sweep over learning-rate and batch-size configurations. Right: extended training progression under the selected configuration, with the best validation action MSE at 30K steps.
Experiments
01
Long-Horizon Tasks
Table: Evaluation of end-to-end success rates on long-horizon tasks.
Task name
π0.5
Motus
FastWAM
GigaWorld-Policy
GigaWorld-Policy-0.5
Food Heating
0.50
0.60
0.50
0.60
0.80
Tableware Arrangement
0.70
0.60
0.50
0.60
0.80
Average
0.60
0.60
0.50
0.60
0.80
Demos
0:00 / 0:00
Heating Food
0:00 / 0:00
Tableware Arrangement
02
Text-Following Tasks
Table: Evaluation of text-following ability on the fruit-picking task.
Text instruction
π0.5
Motus
FastWAM
GigaWorld-Policy
GigaWorld-Policy-0.5
Pick the banana and place it into the basket.
0.88
0.93
0.83
0.90
0.95
Pick the apple and place it into the basket.
0.73
0.78
0.75
0.73
0.80
Pick the lemon and place it into the basket.
0.68
0.75
0.73
0.78
0.83
Pick the grape and place it into the basket.
0.85
0.83
0.88
0.88
0.93
Pick the avocado and place it into the basket.
0.68
0.70
0.75
0.73
0.78
Pick the strawberry and place it into the basket.
0.78
0.80
0.73
0.78
0.85
Average
0.76
0.80
0.78
0.80
0.85
Table: Evaluation of text-following ability on the object-placement task.
Text instruction
π0.5
Motus
FastWAM
GigaWorld-Policy
GigaWorld-Policy-0.5
Pick up the bowl and place it on the plate.
0.87
0.88
0.73
0.80
0.93
Pick up the fork and place it on the plate.
0.78
0.83
0.85
0.80
0.88
Pick up the spoon and place it on the plate.
0.73
0.83
0.80
0.75
0.85
Pick up the bowl and place it into the basket.
0.75
0.83
0.80
0.88
0.95
Pick up the fork and place it into the basket.
0.65
0.75
0.68
0.78
0.85
Pick up the spoon and place it into the basket.
0.80
0.85
0.75
0.83
0.88
Average
0.76
0.83
0.77
0.81
0.89
Citation
@article{gigaworld-policy-0.5,
title={GigaWorld-Policy-0.5: A Faster and Stronger WAM Empowered by AutoResearch},
author={Team, GigaWorld and Ye, Angen and Ma, Angyuan and Wang, Boyuan and Ni, Chaojun and Ye, Fangzheng and Huang, Guan and Li, Guo and Zhao, Guosheng and Yan, Haodong and others},
journal={arXiv preprint arXiv:2607.13960},
year={2026}
}
